Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
It’s been a big week in AI, with some of the people building it calling for a slower pace.
Jacob Coxon’s resignation from Anthropic went viral after he warned that the race towards superintelligence was putting lives at risk.
Then Dario Amodei (Anthropic CEO and co-founder) published his case for “pacing the frontier”, arguing that safety work needs time to catch up with rapidly improving AI capabilities.
Sam Altman and Elon Musk both voiced agreement, a rare moment of alignment between fierce competitors.
Whether that agreement translates into meaningful action is the question to watch. I recommend reading Dario’s essay in full this week: it explains why he believes the pace needs to change and sets out concrete steps for making that happen.
1:1 AI Coaching Programme: Space opening in my next four-week programme. Built around your AI stack, from ChatGPT and Claude through to Claude Code, Codex, and your actual workflows rather than a generic curriculum.
Reply if you want the details.
Happy reading ✌️
Did someone forward you this? Sign up
🇳🇿 New Zealand News
Labour would create an Office of AI, an online safety regulator and four conditions on data centres. Chris Hipkins and Reuben Davidson launched the plan on 9 September.
4 min read
Our take: Thankfully, Labour has put AI regulation on the election agenda. A good start but Labour's AI plan says nothing about how its rules would be enforced or what the new office could do. An Office of AI needs a clear remit, staff and powers.
Antioch raised US$32m led by Greylock, its third round in nine months. Founder Harry Mellsop builds simulation software that Amazon uses to test Ring cameras without hiring actors.
5 min read
Our take: Antioch sells testing software. That part of robotics gets little attention and is a critical feature of the next wave at the intersection of the physical world and AI. Three rounds in nine months is a lot of fundraising, and off to a great start partnering with the single largest user and internal operator of logistics robots in the West in Amazon.
Researchers found 69 of 3,000 peer-reviewed papers carried references invented by AI. The University of Auckland team found one paper citing 29 sources, and not one of them was real.
3 min read
Our take: Firms using AI to search their own documents need automatic citation checks, as an invented reference to an internal file can look real if nobody opens it. An internal AI search tool needs to show the passage supporting its answer and check that the source exists. Most AI systems don’t have the correct checks and balances built in to test these risks.
Netsafe logged more than 30 school deepfake reports in the first half of 2026. That beats the 23 recorded across all of 2024 and 2025, and one video passed 150,000 views.
5 min read
Our take: New Zealand law already treats AI-generated images of under-18s as objectionable, carrying up to 10 years for possession and 14 for distribution. One video passed 150,000 views. A student should not have to work out alone how to report it, ask for its removal and get help from their school.
Datacom found 13% of New Zealand businesses have a chief AI officer. Only 22% have any dedicated AI leader, and the share saying AI transformed core operations fell to 4% from 8%.
4 min read
Our take: 81% are still exploring or implementing AI and we’re nearly 4 years into this journey. That's a long time to call it a pilot when this has become how the business operates. Before buying more tools, businesses should be able to name the work that has changed and show whether it is better.
Spark launched a dedicated interconnect linking Auckland data centres from Silverdale to Papakura. New Zealand now has 56 data centres operating, 20 more in the pipeline and about $10 billion of investment expected.
2 min read
Our take: Local hosting does not guarantee legal sovereignty. Businesses considering it still need to check who owns their provider and which laws apply to their data.
22 ChatGPT Agents Built for Every Marketing Job
Most marketers use ChatGPT to do general research and then call it an AI strategy. The ones outperforming them are deploying specialized agents built for specific jobs.
We put together 22 plug-and-play ChatGPT marketing agents that handle the work eating your week, each with built-in instructions and structured outputs ready to go in under 5 minutes.
Subscribe to Marketing Against the Grain and get all 22 free.
Inside you'll find:
Competitive intelligence agent that visits competitor websites and builds detailed comparison matrices automatically
Customer feedback analyzer that ranks improvement opportunities by business impact
Social listening specialist that monitors brand mentions and flags reputation risks before they escalate
Campaign optimization agents that handle attribution analysis and surface what is actually driving results
Your competitors are already running agents like these.
Get 22 ChatGPT Marketing Agents free when you subscribe to Marketing Against the Grain today.
📚️ Mike's Takes From The Week
Helping leaders and teams learn, adapt, and scale with AI.
1️⃣ Leaders stuck on a 2023 aha AI moment should be fired: Too many executives still think AI means rewriting an email or summarising a PDF. They need to see it do work for them.
Read
2️⃣ Cheaper prototypes unlock more demand for engineers, not less: Martijn Verburg at the Aotearoa AI Summit explained why making software development available to more people suits a small country.
Read
3️⃣ Data centres are consuming all of the oxygen: Greg Davidson on New Zealand's opportunities in global AI.
Read
4️⃣ Design the future process, but do not skip understanding the current one: Accenture opened Summit Day 2 by arguing for a focus on the "to be" process instead of mapping today's.
Read
5️⃣ Quote of the Summit: If your chief executive is ignoring AI, John-Daniel Trask has some advice.
Read
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
Anthropic published its fourth threat intelligence report, covering December 2025 to August 2026 across seven harm areas, with one intrusion exfiltrating 26GB. (Sep 10)
The report accused Alibaba, Moonshot and DeepSeek of five distillation campaigns totalling nearly 200 million Claude exchanges. (Sep 10)
Dario Amodei proposed embedded third-party evaluators with company laptops and a contractual right to publish findings Anthropic cannot redact. (Sep 12)
Smart reports launched in beta on Claude Enterprise, covering team usage, cost, where sessions hit friction and repeated patterns worth packaging as skills. (Sep 10)
Nvidia is in talks to invest up to $10 billion in Anthropic's listing, an offering seeking as much as $100 billion. (Sep 11)
OpenAI
Sam Altman told staff a 10% risk of a catastrophic AI outcome is unacceptable, confirming OpenAI has paused some internal training runs. (Sep 11)
Altman ruled out a 2026 IPO listing, calling it an ill-advised moment to go public given safety concerns. (Sep 12)
The Agents API entered public beta, giving developers a managed Codex runtime with session orchestration, context compaction and recovery at no extra fee. (Sep 10)
ChatGPT for Financial Services launched with Daloopa, PitchBook and LSEG data built in, with Morgan Stanley and Evercore as design partners. (Sep 10)
Gemini can now create Docs, Sheets and Slides, draft email, resolve calendar conflicts and write Tasks from inside any Workspace app. (Sep 9)
DeepMind published AlphaGenome Atlas, predicting the effect of all 9 billion possible single-letter human DNA changes in a 1-petabyte dataset. (Sep 8)
Chinese Frontier Labs
DeepSeek released V4.1-Flash under an MIT licence, a 552 billion parameter model activating 8 billion on prefill and 16 billion on decode. (Sep 10)
V4.1-Flash scored 90.9 on GPQA Diamond and 90.6 on Terminal-Bench 2.1, beating DeepSeek's own 1.7 trillion parameter V4-Pro flagship. (Sep 10)
DeepSeek cut API pricing to $0.15 per million input tokens off-peak and $0.60 per million output, at a 1 million token context. (Sep 10)
Moonshot AI is targeting $2 billion in annualised revenue by year end, doubling an August run rate of $1 billion. (Sep 11)
Open-Weight Models
Nex-AGI released Nex-N2.5-Pro, a 397 billion parameter Apache 2.0 agentic model scoring 61.2% on SWE-Bench Pro and 82.7% on Terminal-Bench 2.1. (Sep 8)
NVIDIA open-sourced a maths training recipe that scored 30 of 42 points at the 2026 International Mathematical Olympiad, with weights, data and code. (Sep 9)
Edge0 released a 35 billion parameter mixture-of-experts model that streams experts from SSD and decodes at 15 tokens a second. (Sep 8)
Microsoft
Microsoft added xAI's Grok models to Copilot in Word, Excel and PowerPoint through the Frontier Program, off by default and unavailable in the EU and UK. (Sep 12)
Copilot Business and Enterprise admins gained three central permission levels over agent shell commands, file edits and network domains, which users cannot override. (Sep 9)
Copilot code review moved to an ensemble of review agents, lifting addressed high-severity findings 47% while cutting review cost about 8%. (Sep 11)
NVIDIA
NVIDIA and Palantir launched a joint sovereign AI stack using Nemotron open models, first deployed on NVIDIA's own supply chain of 1.3 million parts. (Sep 10)
Eight Australian data centre providers will build toward 2 gigawatts of NVIDIA capacity by 2027, with Sharon AI deploying up to 68,000 GPUs. (Sep 9)
Amazon, Meta and Mistral
GPT-6 Astra reached general availability on Amazon Bedrock with a context window up to 1 million input tokens. (Sep 8)
SageMaker HyperPod added model caching, giving about 60% faster scale-out and dropping cold starts from tens of minutes to seconds. (Sep 11)
Meta launched Muse, a personal AI agent on iOS, Android and the web, running each user on a dedicated secure virtual machine, in the United States only. (Sep 8)
Mistral raised 3 billion euros led by Samsung Electronics at over 21 billion euros post-money, the largest equity round by a European technology company. (Sep 8)
👋 Mike & Erin